|Publication number||US8046761 B2|
|Application number||US 11/527,325|
|Publication date||Oct 25, 2011|
|Filing date||Sep 26, 2006|
|Priority date||Sep 26, 2005|
|Also published as||EP1932077A2, US20070101013, WO2007034232A2, WO2007034232A3|
|Publication number||11527325, 527325, US 8046761 B2, US 8046761B2, US-B2-8046761, US8046761 B2, US8046761B2|
|Original Assignee||Imagination Technologies Limited|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (33), Non-Patent Citations (8), Referenced by (4), Classifications (12), Legal Events (2)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This invention relates to a system capable of processing multi-media data such video and 3-dimensional computer graphics and in particular to methods of processing multiple parallel streams of such data within a scalable multi-threaded environment.
International patent application number W097/38372 (the contents of which are incorporated herein by reference) describes a system for processing multiple real time data sources through a multi-threaded media processing core. The processing core handles a number of execution of instructions. A determination is made as to which thread has the highest priority for execution and execution is formed on that thread in accordance with this.
International patent application number WO02/06716 entitled to control a priority lead structure on a multi-threaded processor (the contents of which are incorporated herein by reference) elaborates further on the arbitration schemes which are possible between different executing threads. It provides a method apparatus for controlling execution rates for instruction threads. The rate at which instructions are to be executed is stored and requests are issued to cause instruction to execute in response to the stored rate. The stored rate is reduced in response to execution of instructions and increased in the absence of instructions for execution. Furthermore, instruction rate is controlled by storing the average rate at which each thread should execute instructions. A value representing the number of instructions available not yet issued is monitored and decreased in response to instruction execution. Also, a ranking order can be assigned to a plurality of instruction threads using a plurality of metrics relating to the threads.
In a further development, a technique known as ‘Enhanced interleaved multi-threading’ works by maintaining a superset of execution threads from which a subset is maintained a active. The number of active threads is determined by the latency of the execution unit to which they are issuing instructions. Threads are then swapped between the active subset and the superset based on resource dependencies, i.e., if an active thread is waiting or the data to be returned in memory it will be swapped for a currently inactive one that is ready to proceed.
Preferred embodiments of the present invention provide a system which enables a muti-threaded processing pipeline to be used to processes multiple input data streams that represent bulk non real time data sources in a manner that maximises latency absorption and utilization of pipeline processing resource by using a multi-phased approach to scheduling.
This is performed by using a three phase scheduling system which differentiates the management of data storage within the promising pipeline from the allocation of proceed to within the same pipeline and the scheduling of instructions based on execution resource conflicts.
In accordance with an embodiment of a first aspect of the invention there is provided a method for processing multiple streams of data on a plurality of execution threads comprising the steps of selecting data from at least one of a plurality of data sources in dependence on the availability of storage in a data storage means, allocating an address in the data storage means for the data thus selected, loading selected data at the data address thus allocated, constructing an execution task comprising the selected data source, the data address, and an execution address, queuing the task with other previously constructed tasks, determining which processing resources are required for each task, selecting tasks for execution in dependence on the processing resources available, distributing tasks selected for execution across a plurality of processing threads, determining whether any of the threads will stall because of data dependencies, repeatedly selecting a subset of threads to be active that will not stall due to said data dependencies, and, extecutng active threads.
Preferably an embodiment includes a means for scheduling tasks based on available resource.
Preferably an embodiment includes a means for avoiding prolonged blocking in a data source.
Preferably an embodiment includes a method for loading data into a unified storage area by way of a programmable data sequencer.
Preferably an embodiment may be modified such that the programmable data sequencer is provided on a data processing thread.
Preferably an embodiment provides a method for ensuring data is processed in the correct order.
Preferably the method may be embodied in a tile based rendering system.
Preferably the programmable data sequencer is used for texture coordinate iteration and texture fetches through the system.
Preferably embodiments of the invention are scalable by the different multi processors cores (wherein each multi processor core is matched to a different sector of a tile, the tile size may be adjusted based on the number of attached pipelines, the vertex data is distributed evenly across the pipelines using a pre-scheduling block, and a vent processor is distributed evenly across the pipelines also by a pre-scheduling block).
In accordance with an embodiment of a further aspect of the invention there is provided, a method for allocating data storage in a multithreaded data processing system comprising the steps of selecting data from one of a plurality of data so sources in dependence on the availability of data storage, supplying a code execution address to a programmable data sequencer, executing code from the code execution address to cause data to be written to the data storage means.
Preferred embodiments of the invention will now be described in detail by way of example with reference to the accompanying drawings in which:
FIG. 1—illustrates the proposed scheduling phases of an embodiment of the invention discussed above;
FIG. 2—illustrates the Improved Multi-Media Processing Architecture embodying the invention;
FIG. 3—detailed diagram of architecture of
FIG. 4—illustrates thread States within Thread Manager of
FIG. 5—illustrates a typical Tile Based Rendering System;
FIG. 6—illustrates a typical HSR unit within a tile based rendering system which may use an embodiment of the invention;
FIG. 7—illustrates a tile based rendering system implemented using an embodiment of the invention;
FIG. 8—illustrates tile sector to Pixel Data Master (PDM) mapping;
FIG. 9—illustrates a system scaled to two pipelines;
The second scheduling phase involves each task within the queue generated by the first phase being distributed across one or more processing threads within a multi-threaded processing pipeline based on the availability of processing resource within that pipeline. The processing resources required to execute any given task are used to dictate which task within the task queue may be run at 162, specifically the required number of processing threads and temporary storage required to execute those threads. The selected task is to distributed across execution threads 170 at 164.
The third and final execution phase determines which threads may be run at 175 without incurring a stall due to resources that are not currently available e.g. data returned by an external memory read or branch condition that is not currently valid. This is done in a manner that combines the simple resource checking described in patent WO 907/38372 and the technique of “Enchanced interleaved multi-thread”. This results in a maintained set of active 180, the number of which is tied to the execution latency of “Execute” 190 (typically an ALU). It should be noted that the third phase executes threads to completion at which point both the threads and any allocated memory resource are returned to the system.
The DMS cycles through the data masters looking for one that may be serviced within the resources that are currently indicated as free by a Resource Manager 1002 as described in scheduling phase 1 above. When selecting a data master the DMS also takes into account a minimum data instance count (Referred to as the Data Master Minimum Service Count) that is indicated by the data masters. There must be sufficient resources available to service a multiple of the minimum data instance count stipulated. Once a data master has been selected the DMS allocates the required resources updating the Resource manager as appropriate before passing the selected data master's request onto a Programmable Data Sequencer (PDS) 1004. It should be noted that the resource allocation is in the form of input and output data storage within the MTMP's shared data store 1036.
Input buffer storage is allocated on a per service request basis, however output buffer storage needs additional management in certain circumstances. Specifically certain data types require output buffering to be persistent across multiple service requests (e.g. pixel data), in these cases the output buffering is managed in terms of “macro” tasks. A macro task being defined as a group of task over which output buffering must be persistent. To handle this the DMS/resource manager allocates output buffer storage only on the first task for a data master where there is no output buffer storage currently allocated. This allocation then remains persistent until the data master signals that the macro task is complete, where it may be freed.
It will be noted that it is possible for a data master that requires a large amount of resource for its minimum serviceable quanta to be blocked from processing. This is prevented by the DMS maintaining a timeout counter for each data master. This timeout counter is incremented each time it has a service request rejected, when this counter reaches a pre-determined value the data master concerned will be serviced irrespective of currently available resources i.e. the DMS will be prevented from processing until sufficient resources are available. In addition a software control is made available to control and disable/enable the servicing of each data master. This software control can be used to manipulate the amount of processing time available to each data master.
The PDS is a fully programmable processor that is supplied with a code execution address by a data master selected by the DMS. The indicated code is executed once for each data instance that is being serviced from the data master. The code may use data supplied by the data master to generate data directly into the shared data store 1036 or to fetch data from memory via a DMA unit 1008 into the shared data store 1036. It should be noted that the PDS could be implemented as a thread within the MTMP. However this would then take processing resource from other activities.
Once all data instances from a selected data master have been serviced by the PDS a task is issued to the MTMP's task queue 1022. The execution address for the task is preferably supplied directly from the code executed within the PDS. Thus the PDS operation is equivalent to the “Data Load” step 130 from
The task queue is scanned by a Task Control Unit (TCU) 1024 for tasks that can be issued to a Thread Manager 1026, this is any task that,
Execution dependencies are used to block the execution of a task while there is some external or internal activity that must complete before the task can be allowed to proceed. Examples of such depending are (but not limited to an “external” dependency or a “sequential” dependency.
An external dependency is a condition which is sourced from a piece of hardware external to the MTMP which blocks the task's execution until the condition is released. For example a task may require one of the processing pipelines 1040 to complete before proceeding.
A sequential dependency is applied where a task may update an output buffer register that is also updated on a task that preceded its arrival in the queue. When the task manager sees a task with a sequential dependency set it will allow all tasks from the same data master in the queue that were supplied before it to complete before allowing it to proceed. It should be noted that only tasks from the same data master are blocked in these circumstances, this is why the data master/source ID is included within the task data. In a further modification to this mechanism a task could carry a mask of output registers that are modified, the sequential dependency then being applied selectively only where those masks overlap.
In order to prevent a task with large resource requirements from being blocked for long periods of time each task in the queue has an “Age” counter which is incremented each time a task is rejected for processing by the TCU. When the age counter reaches a defined limit a task that has no outstanding execution dependencies will be selected for processing irrespective of available thread and temporary storage availability.
Once a task has been selected for service its data instances are distributed to threads as they and the required temporary register resources become available as shown schematically in the scheduling phase 2 in
The Thread Manager 1026 performs the actions to implement scheduling phase 3 as shown in
The thread manager issues instruction fetches for active threads to the Instruction Fetch and Decode unit (IFD) 1028 which issues fetches to the instruction cache 1030. If an issued fetch causes a cache miss the IFD informs the thread manager of this, which will deactivate the thread as described above, until such time as the Instruction cache retrieves the required cache line. Fetched instructions are decoded and any branch or control instruction executed within the IFD. Any control or branch instruction that references a control resource (such as a condition code) that is not currently valid will result in the IFD informing the thread manager that the thread should be deactivated until such point as the control resource becomes valid. All other instructions are passed to the ALU 1034 for execution. It will be noted the ALU operates directly on the data contained within the shared data store.
Prior to completion most threads will signal an emit to the Processing Pipelines 1040 which will read required data from the output buffers allocated within the shared data store 1036. When this process is complete the processing pipeline signals that the output buffer resource has been freed.
In a preferred embodiment of the invention the system is adapted for use within a tile base rendering system.
Tile based rendering systems are known. These break down an image to be rendered into a plurality of rectangular blocks or tiles. The way in which this is done and the subsequent texturing and shading performed is shown schematically in
The tiled data lists are fetch by a tile parameter fetch unit 1505 which feeds them tile by tile to a hidden surface removal unit (HSR) 1506 and from there to a texturing and shading unit (TSU) 1508. The HSR unit processes each primitive in the tile and passes only data for visible pixels to the TSU as described below.
The texturing and shading unit takes the dam from the HSR unit and uses it to fetch textures and apply shading to each pixel within a visible object using techniques that are well known.
The TSU then feeds the textured and shaded data to the alpha test unit 1509, fogging unit 1510 and alpha blending unit 1511 which perform alpha test, fogging and alpha blending respectively in a well known manner. Alpha blending is performed to an on chip tile buffer 1512 thereby eliminating external memory bandwidth associated with this operation. On completion of each tile the Pixel processing unit 1514 performs any backend processes such as packing or anti-alias filtering before writing the resulting data out to external memory used to store the rendered scene 1516.
In this example the MTMP replaces the function of geometry processing 1502, texturing and shading 1508, alpha test 1509, fogging 1510 mad alpha blend 1511 unit from
The Data masters 1820 are now allocated specific functions as follows. The pixel data master (PDM) 1826 incorporates the functionality of tiled parameter fetch 1505 and hidden surface removal 1506 in
The vertex data master (VDM) 1822 fetches incoming primitives from external memory and passes indices for vertices that should be fetched in each data service request. It should be noted that the VDM also directly passes sideband information to the tiling engine 1862 to enable it to correctly re-construct primitives from the processed vertices.
The event data master (EDM) 1824 is fed with events from the other data masters within the system and from the Host allowing the architecture to be used for processing system events and general commands avoiding the need to interrupt the host processor to service these events.
It should be noted that additional data masters can be added to the system, for example a data master could be implemented that takes a feed from the tiled parameter fetch unit allowing the fetched primitive data to be pre-processed before it is passed to the HSR unit (which would in effect become another processing pipeline). This pre-processing would allow for generation of procedural geometry such as high order surfaces while the tiled lists are being processed allowing for a significant reduction in consumed memory for tiled primitive lists.
The processing pipelines 1860 are also now allocated specific functions with the tiling engine 1862, the pixel processing 1864 and the texturing unit 1866 being directly equivalent to same units in
The approach to data processing is a described previously with a few additional functions available. Specifically the PDS may now, for pixel data, also load data into the shared data store by way of a set of coordinate iterators 1810 or the texturing unit 1866.
Pixel processing occurs whenever the PDM's HSR unit spawns a pass which will cause it to signal the DMS that it has data to be processed. When selected the code executed on the PDS loads coordinate and texture data into the allocated area of the shared data store. Data instances within the system represent a single pixel with executed code storing the results into the allocated output buffer within the shared data store. It should be noted that each pixel data instance does not perform an emit to the pixel processing unit 1864 as the output buffer is not finalised until the entire tile has be processed. In order for the tile data to be emitted the EDM 1824 is fed by a signal from the PDM that indicates that the tile is complete. The EDM then generates a task that is used to emit the tiled data to the pixel processing it. It should be noted that the EDM can be used to generate events in this manner for many other purposes, for example a task could be emitted between passes spawned by the HSR unit allowing processing to be performed on the contents of the tile buffer between each pass. It should be noted at it is possible to perform any amount or type of processing on these events, for example when emitting the completed tile additional processing could be performed to generate levels for a mip-map (a well known technique) for the emitted data.
Vertex processing operates in a similar manner to a generic data source as described previously with an emit being performed to the tiling engine per processed vertex.
In a further extension to the embodiment the architecture may also be scaled. For pixel processing this is accomplished by laying down multiple media processing cores with each core mapped to a different “sector” of the tile as illustrated in
As these different sector are non overlapping they can inherently be processed in parallel across multiple pipelines. Typically as the number of pipelines are increased the tile size is increased in proportion to the number of attached pipelines e.g. if a single pipeline is attached to an HSR unit that works on a 16×16 tile, then two pipelines would be attached to an HSR unit that works on a 32×16 tile and so on. Alternatively the size of the sectors can be reduced allowing the tile size to be kept constant, however this will result in a reduction of efficiency when processing pixel data in each pipeline.
Generally speaking the approach of scaling the tile size to match the number of pipelines can continue until the performance of the HSR unit is saturated i.e. it becomes difficult to scale its performance further. At this point scaling can be continued by laying down multiple parallel HSR units.
For vertex processing scalability can be achieved by distributing the incoming primitive data across the available pipelines. This is typically done using either a simple round robin mechanism or by sending blocks of vertices to the pipeline which is the least busy. The former mechanism is preferred as it simplifies the tiling mechanism later in the pipeline.
The HSR unit 2200 processes incoming tiled parameter lists as previously described. The pixel data masters 2230 and 2231 are each mapped to a different part of the HSR unit's tag buffer as described above. Each PDM is dedicated to a specific media processing core 2250, 2251. It should be noted that the texturing (2260,2261) and pixel processing (2270, 2271) pipelines arm duplicated along with the media processing core. The processed pixel data is passed from the MTMP's to the pixel processing units 2270, 2271 at the completion of a tile as previously described. The pixel processing units are duplicated along with the processing units as they have a non overlapping assignment to pixels stored in memory making them directly scalable along with the rest of the pixel processing pipeline.
The vertex data master 210 and the event data master 2220 each feed service requests into the pre-scheduler 2240. The pre- scheduler maintains a service queue per attached pipeline for each of data masters (VDM, EDM). The queues are filled in simple round robin order a processing requests are distributed evenly across the attached pipelines. Processed vertices are passed from the MTMP to the tiling unit 2280 which re-combines them with primitive data generated by the VDM before applying the tiling process as previously described.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5339413 *||Aug 21, 1992||Aug 16, 1994||International Business Machines Corporation||Data stream protocol for multimedia data streaming data processing system|
|US5412788||Apr 16, 1992||May 2, 1995||Digital Equipment Corporation||Memory bank management and arbitration in multiprocessor computer system|
|US5517631||Jul 7, 1993||May 14, 1996||Quantum Corporation||Miniature disk drive having embedded sector servo with split data fields and automatic on-the-fly data block sequencing|
|US5860018 *||Jun 25, 1997||Jan 12, 1999||Sun Microsystems, Inc.||Method for tracking pipeline resources in a superscalar processor|
|US5968167 *||Apr 3, 1997||Oct 19, 1999||Videologic Limited||Multi-threaded data processing management system|
|US6359623||Nov 12, 1998||Mar 19, 2002||Hewlett-Packard Company||Method and apparatus for performing scan conversion in a computer graphics display system|
|US6415340||Sep 15, 2000||Jul 2, 2002||Micron Technology, Inc.||Method and apparatus for synchronous data transfers in a memory device with selectable data or address paths|
|US6418458||Oct 2, 1998||Jul 9, 2002||Ncr Corporation||Object-oriented prioritized work thread pool|
|US6523098||Dec 22, 1999||Feb 18, 2003||Intel Corporation||Mechanism for efficient low priority write draining|
|US6675190 *||Oct 7, 1999||Jan 6, 2004||Alcatel||Method for cooperative multitasking in a communications network, and a network element for carrying out the method|
|US6728862||May 22, 2000||Apr 27, 2004||Gazelle Technology Corporation||Processor array and parallel data processing methods|
|US6940514||Apr 26, 2004||Sep 6, 2005||Sun Microsystems, Inc.||Parallel initialization path for rasterization engine|
|US6948050||Jun 6, 2001||Sep 20, 2005||Texas Instruments Incorporated||Single integrated circuit embodying a dual heterogenous processors with separate instruction handling hardware|
|US7080377 *||Jun 29, 2001||Jul 18, 2006||Eci Telecom Ltd.||Method for effective utilizing of shared resources in computerized system|
|US20020087840||Dec 29, 2000||Jul 4, 2002||Sailesh Kottapalli||Method for converting pipeline stalls to pipeline flushes in a multithreaded processor|
|US20050021930||Jul 9, 2003||Jan 27, 2005||Via Technologies, Inc||Dynamic instruction dependency monitor and control system|
|US20050021931||Feb 19, 2002||Jan 27, 2005||Anderson Adrian John||Control of priority and instruction rates on a multithreaded processor|
|US20050081195 *||Oct 14, 2003||Apr 14, 2005||Shailender Chaudhry||Selectively deferring the execution of instructions with unresolved data dependencies as they are issued in program order|
|US20060179280||Feb 4, 2005||Aug 10, 2006||Mips Technologies, Inc.||Multithreading processor including thread scheduler based on instruction stall likelihood prediction|
|US20060190703||Feb 24, 2005||Aug 24, 2006||Microsoft Corporation||Programmable delayed dispatch in a multi-threaded pipeline|
|EP0697654A1||Jul 17, 1995||Feb 21, 1996||Canon Kabushiki Kaisha||Load distribution method and system|
|EP0725367A1 *||Jan 30, 1996||Aug 7, 1996||Videologic Limited||Improvements relating to computer 3D rendering systems|
|EP0843262A2||Nov 11, 1997||May 20, 1998||Hyundai Electronics America, Inc.||Memory request reordering in a data processing system|
|EP0862113A2||Feb 27, 1998||Sep 2, 1998||Zuno Limited||Autonomous agent architecture|
|EP1696317A2||Jan 16, 2006||Aug 30, 2006||Microsoft Corporation||Programmable delayed dispatch in a multi-threaded pipeline|
|GB2311882A *||Title not available|
|GB2372349A||Title not available|
|WO1997038372A1||Apr 4, 1997||Oct 16, 1997||Videologic Ltd||A data processing management system|
|WO1999019805A1||Oct 9, 1998||Apr 22, 1999||Rambus Inc||Method and apparatus for two step memory write operations|
|WO2000033186A1||Dec 3, 1999||Jun 8, 2000||Sun Microsystems Inc||Variable issue-width vliw processor|
|WO2001090915A2||May 22, 2001||Nov 29, 2001||Gazelle Technology Corp||Processor array and parallel data processing methods|
|WO2002067116A2||Feb 19, 2002||Aug 29, 2002||Imagination Tech Ltd||Control of priority and instruction rates on a multithreaded processor|
|WO2006083543A2||Jan 18, 2006||Aug 10, 2006||Mips Tech Inc||Multithreading processor including thread scheduler based on instruction stall likelihood prediction|
|1||"Enhanced Interleaved Multithreaded Multiprocessors and Their Performance Analysis", by Zuberek, Institution of Electrical Engineers, GB, 2004.|
|2||"Implicitly-multithreaded Processors", by Park et al, Proceedings of the 30th Annual Int'l Symposium on Computer Architecture (ISCA '03), USA, Jun. 2003.|
|3||"Performance Analysis of Enhanced Fine-Grain Multithreaded Distributed-Memory Systems", by Zuberek, IEEE International Conference on Systems Man and Cybernetics, USA, Oct. 2001, pp. 1101-1106.|
|4||2004 International Conference on Dependable Systems and Networks, N. Nakka and Z. Kalbarczyk, "An Architectural Framework for Providing Reliability and Security Support" (10 sheets).|
|5||Proceedings of the Fourth International Conference on Application to Concurrency to System Design, W.M. Zuberek, "Enhanced Interleaved Multithreaded Multiprocessors and Their Performance Analysis" (0-7695-2077-4/2004).|
|6||U.K. Search Report dated Dec. 21, 2006 (1 sheet).|
|7||United Kingdom Patent Office Search Report dated Nov. 30, 2007 (6 pages).|
|8||United Kingdom Search Report dated Jan. 26, 2006 (1 page).|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US8669987||Feb 15, 2007||Mar 11, 2014||Imagination Technologies, Ltd.||Memory management for systems for generating 3-dimensional computer images|
|US8752061 *||Nov 24, 2008||Jun 10, 2014||Freescale Seimconductor, Inc.||Resource allocation within multiple resource providers based on the incoming resource request and the expected state transition of the resource requesting application, and selecting a resource provider based on the sum of the percentage of the resource provider currently used, the requesting load as a percentage of the resource provider's total resource, and the additional load imposed by the expected state of the requesting application as a percentage of the resource provider's total resource|
|US20080186318 *||Feb 15, 2007||Aug 7, 2008||Jonathan Redshaw||Memory management for systems for generating 3-dimensional computer images|
|US20110214129 *||Nov 24, 2008||Sep 1, 2011||Freescale Semiconductor, Inc.||Management of multiple resource providers|
|U.S. Classification||718/102, 712/225|
|International Classification||G06F9/46, G06F9/38, G06F9/00|
|Cooperative Classification||G06F9/5011, G06F2209/507, G06F9/3824, G06F9/3851|
|European Classification||G06F9/50A2, G06F9/38D, G06F9/38E4|
|Dec 18, 2006||AS||Assignment|
Owner name: IMAGINATION TECHNOLOGIES LIMITED, UNITED KINGDOM
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:HOWSON, JOHN;REEL/FRAME:018646/0234
Effective date: 20061129
|Apr 16, 2015||FPAY||Fee payment|
Year of fee payment: 4